feat: make the workspace handle reachable — provision furrow, fix the skill, expose the address

[AbirAbbas]

Summary

Testing #885 + Agent-Field/SWE-AF#130 against a live cloud deployment showed the workspace-sync feature was unreachable end to end, for three independent reasons. This fixes all three, plus a red test on main found on the way.

A workspace_handle is only usable if furrow is on the caller's machine, and furrow had no distribution channel — so the honest instruction was "install Rust and compile it." Agent-Field/furrow now publishes releases (Agent-Field/furrow#6), so af can provision the binary like everything else under ~/.agentfield/.

Provisioning lives in the install path deliberately. The alternative — teaching the skill to curl it at use time — puts an installer inside a document, with no checksum, no version pin, and no single place to fix it.

Changes Made

Five focused commits:

1. fix(skills): sync the embedded agentfield-use mirror. feat(desktop): open the workspace-sync port when we deploy a control plane #885 added the workspace-handle section to skills/ but not to the embedded copy the control plane serves. Three internal/skillkit tests have been failing on main since that merge; this is scripts/sync-embedded-skills.sh and nothing else.
2. feat(control-plane): expose the furrow address in health. The desktop probe from feat(desktop): open the workspace-sync port when we deploy a control plane #885 (furrowAddressFrom in cloud.ts) reads furrow_public_addr from the health body, but nothing emitted it, so furrowAvailable could never be true. Emitted from the shared handler when FURROW_PUBLIC_ADDR is set; the key is absent when it is not.
3. feat(af): provision the pinned furrow client. Downloads the pinned release asset to ~/.agentfield/bin/furrow, verified against the release's SHA256SUMS, temp-file-and-rename so a failure never leaves a partial binary, size-bounded responses. Unsupported platforms are a clean no-op — there is no Windows asset because furrow uses std::os::unix unconditionally in 13 files, and Windows users reach it through WSL. The installed version is recorded beside the binary so bumping the pin actually upgrades machines that already have furrow.
4. feat(af): install it with the skill that uses it, best-effort so a failed download never fails the install, plus af furrow ensure for explicit repair — which does return the error, since silence is right for an install that merely offers to provision furrow and wrong for someone who asked by name.
5. docs(skill): resolve furrow from where AgentField installs it. The skill said to use the handle "only if furrow is on PATH" — which, with no release channel, silently disabled the feature for every caller. Now a pure resolver over ~/.agentfield/bin and a node's vendored copy, keeping the silent skip when neither exists.

Test Plan

• go test ./internal/furrow ./internal/skillkit ./internal/server — pass (skillkit is red on main before commit 1)
• Behaviors covered: platform→asset mapping incl. the Windows no-op, idempotent skip, upgrade when the pin changes, marker recorded so the next run skips, checksum mismatch rejected with nothing written, download failure non-fatal, rename failure leaves no partial file, opt-out honored, health key present/absent
• go build ./..., go vet ./..., gofmt clean
• No network in any test (httptest via a base-URL override)
• The resolver was run verbatim on a real machine against an already-vendored binary
• End-to-end clone of a live cloud run's workspace verified against a Railway deployment

Supersedes #888 and #889. Pairs with Agent-Field/SWE-AF#130.

Review hardening (second pass)

A five-dimension review swarm over the full upgrade story (provisioning diff, trigger coverage, desktop upgrade path, cloud image, release pipeline) produced five more commits:

6. fix(furrow): serialize concurrent installs, fix download timeouts. Two processes ensuring at once (worse, two different af versions) could interleave rename/marker and permanently mark a stale binary current — an flock around the sequence serializes them, with a re-check under the lock. The 15s whole-request timeout failed spuriously on slow links for a ~7.5MB asset; replaced with phase timeouts (dial/TLS/header) under a 3-minute ceiling. Race-clean concurrent test asserts exactly one download.
7. fix(skill): POSIX resolver, AGENTFIELD_HOME, version bump. The resolver used bash-only {bin,go/bin} brace expansion (dash leaves it literal — furrow-dial never found under sh) and hardcoded ~/.agentfield while provisioning honors AGENTFIELD_HOME. The catalog demands a version bump on every content change; agentfield-use is now 0.5.0 so reconcilers have a real signal.
8. fix(deploy): AGENTFIELD_SKIP_FURROW=1 in the cloud image. The client is a laptop-side tool; cloud agents get furrowd vendored by their packages. Nothing in the container should pull it onto the volume.
9. feat(desktop): visible cloud upgrades + skill sync in cloud mode. Re-run deploy has been a safe upgrade since image pinning, but nothing said an upgrade existed: the panel now compares the deployed pin against Docker Hub and shows "Control plane vX → vY available" with an "Upgrade & redeploy" action, and the connection test gains a Workspace-sync row (neutral for servers predating the health field). Also removes the !isCloudActive() guard on syncSkills — skills and the furrow client belong to local coding agents regardless of where the control plane runs, and a cloud-connected laptop is exactly the machine that needs the workspace client.
10. chore(furrow): errcheck-clean provisioning code.

Gates run locally: full control-plane suite (UI build → go build → go test ./...; the only failure is TestDevServiceRunDev, which fails identically on origin/main in this environment — the known WSL port-discovery flake), golangci-lint --new-from-rev origin/main = 0 issues, desktop npm ci + typecheck + vitest (402 passing) + dist:dir, cloud image built + smoke-tested including the new env.

Known follow-ups deliberately not in this PR: release.yml publishes binaries before the cloud image (split-release window if the image push fails); furrow-dial still has no standalone channel (fold into af later); a bare binary-swap upgrade triggers no skill reconcile (curl installer and desktop are the covered paths); no linux-arm64 furrow asset yet.

---

🤖 Generated with Claude Code

[github-actions]

📊 Coverage gate

Thresholds from .coverage-gate.toml: per-surface ≥ 84%, aggregate ≥ 85%, max per-surface regression ≤ 1.0 pp, max aggregate regression ≤ 0.50 pp.

Surface	Current	Baseline	Δ
control-plane	87.10%	87.40%	↓ -0.30 pp	🟡
sdk-go	92.70%	92.00%	↑ +0.70 pp	🟢
sdk-python	93.82%	93.73%	↑ +0.09 pp	🟢
sdk-typescript	91.17%	90.42%	↑ +0.75 pp	🟢
web-ui	84.75%	84.79%	↓ -0.04 pp	🟡
aggregate	85.62%	85.75%	↓ -0.13 pp	🟡

✅ Gate passed

No surface regressed past the allowed threshold and the aggregate stayed above the floor.

[github-actions]

📐 Patch coverage gate

Threshold: 80% on lines this PR touches vs origin/main (from .coverage-gate.toml:thresholds.min_patch).

Surface	Touched lines	Patch coverage	Status
control-plane	179	89.00%	✅
sdk-go	0	—	➖ no changes
sdk-python	0	—	➖ no changes
sdk-typescript	0	—	➖ no changes
web-ui	0	—	➖ no changes

✅ Patch gate passed

Every surface whose lines were touched by this PR has patch coverage at or above the threshold.